Real-time structural condition assessment based on multi-source SHM data using attention-based graph convolutional networks

Structural condition assessments based on comprehensive utilization of multi-source real-time monitoring data are helpful for revealing the evolutionary law of structural characteristics, which provides a significant basis for developing maintenance strategies. However, it is challenging to explore complicated correlations among monitoring data under time-varying environmental and operational effects. This paper proposed a real-time structural condition assessment method based on monitoring data forecasting model using multi-source data. The structural response at target time was predicted by historical observation of structural response and current environmental/structural temperature. Attention-based graph convolutional networks and 1-dimensional dilated convolutional networks were used for spatiotemporal modeling of structural response in the forecasting model. The temperature data at target time was also introduced in the prediction using fully-connected networks. According to the Mahalanobis distance of the model residuals, structural anomaly can be detected timely if any deviation of real-time data pattern appears. A case study on a long-span cable-stayed bridge with multiple monitoring items was carried out to verify the effectiveness of the proposed methodology in real-time condition assessment.